AI/rocm-systems
2
0
Derivar 0
Código Questões Pedidos de integração Operações Pacotes Planeamentos Lançamentos Wiki Trabalho

Update to use rocprofiler-sdk (#55)

Ver a proveniência 
- Renames the CMake option "ROCPROFSYS_USE_HIP" to "ROCPROFSYS_USE_ROCM"
- Remove the "ROCPROFSYS_USE_ROCM_SMI option. Controlled with the "ROCPROFSYS_USE_ROCM" option, instead.
   - Runtime configuration can still toggle ROCPROFSYS_USE_ROCM_SMI to disable the sampling.
- Rename ROCPROFSYS_HIP_VERSION macro to ROCPROFSYS_ROCM_VERSION and remove blocks for `ROCPROFSYS_ROCM_VERSION < 60000`
- Remove ROCPROFSYS_USE_ROCTRACER and ROCPROFSYS_USE_ROCPROFILER
- Update test cases
- Update docker files and workflows to install cmake 3.21, which is required for the rocprofiler-sdk findPackage script.
- Removed rocm-6.2 from workflows due to a rocprofiler-sdk API change.
Este cometimento está contido em:
David Galiffi
2024-12-13 18:48:39 -05:00
cometido por GitHub
ascendente d3725df816
cometimento 88aa2d3cbe
87 ficheiros modificados com 3842 adições e 6261 eliminações
Descarregar ficheiro patch Descarregar ficheiro diff Expandir todos os ficheiros Colapsar todos os ficheiros
source/bin/CMakeLists.txt
+3 -12
Ver ficheiro
@@ -1,17 +1,8 @@
# executable RPATH
if(ROCPROFSYS_USE_ROCPROFILER
AND rocprofiler_LIBRARY_DIR
AND ROCmVersion_TRIPLE_VERSION VERSION_LESS 5.2.0
AND NOT CMAKE_INSTALL_RPATH_USE_LINK_PATH)
set(ROCPROFSYS_EXE_INSTALL_RPATH
"\$ORIGIN/../${CMAKE_INSTALL_LIBDIR}:\$ORIGIN/../${CMAKE_INSTALL_LIBDIR}/${PROJECT_NAME}:${rocprofiler_LIBRARY_DIR}"
)
else()
set(ROCPROFSYS_EXE_INSTALL_RPATH
"\$ORIGIN/../${CMAKE_INSTALL_LIBDIR}:\$ORIGIN/../${CMAKE_INSTALL_LIBDIR}/${PROJECT_NAME}"
)
endif()
set(ROCPROFSYS_EXE_INSTALL_RPATH
"\$ORIGIN/../${CMAKE_INSTALL_LIBDIR}:\$ORIGIN/../${CMAKE_INSTALL_LIBDIR}/${PROJECT_NAME}"
)
# executables
add_subdirectory(rocprof-sys-avail)
source/bin/rocprof-sys-avail/avail.cpp
+11 -10
Ver ficheiro
@@ -33,8 +33,7 @@
#include "api.hpp"
#include "core/config.hpp"
#include "core/gpu.hpp"
#include "core/hip_runtime.hpp"
#include "library/rocprofiler.hpp"
#include "library/rocm.hpp"
#include <timemory/components.hpp>
#include <timemory/components/definition.hpp>
@@ -119,7 +118,7 @@ write_hw_counter_info(std::ostream&, const array_t<bool, N>& = {},
namespace
{
// initialize HIP before main so that librocprof-sys is not HSA_TOOLS_LIB
int gpu_count = rocprofsys::gpu::hip_device_count();
int gpu_count = rocprofsys::gpu::device_count();
// statically allocated shared_ptrs to prevent use after free errors
auto timemory_manager = tim::manager::master_instance();
@@ -508,15 +507,15 @@ main(int argc, char** argv)
return EXIT_FAILURE;
}
#if ROCPROFSYS_USE_HIP > 0
#if ROCPROFSYS_USE_ROCM > 0
if(gpu_count > 0)
{
size_t _num_metrics = 0;
try
{
// call to rocm_metrics() will add choices to ROCPROFSYS_ROCM_EVENTS setting
// call to rocm_events() will add choices to ROCPROFSYS_ROCM_EVENTS setting
// so always perform this call even if list of HW counters is not requested
_num_metrics = rocprofsys::rocprofiler::rocm_metrics().size();
_num_metrics = rocprofsys::rocm::rocm_events().size();
} catch(std::runtime_error& _e)
{
verbprintf(0, "Retrieving the GPU HW counters failed: %s", _e.what());
@@ -615,9 +614,9 @@ main(int argc, char** argv)
}
}
signal(SIGABRT, &dump_log_abort);
signal(SIGSEGV, &dump_log_abort);
signal(SIGQUIT, &dump_log_abort);
// signal(SIGABRT, &dump_log_abort);
// signal(SIGSEGV, &dump_log_abort);
// signal(SIGQUIT, &dump_log_abort);
if(!os) os = &std::cout;
@@ -641,6 +640,8 @@ main(int argc, char** argv)
}
dump_log();
const_cast<std::shared_ptr<tim::settings>&>(tim::settings::shared_instance()).reset();
return 0;
}
@@ -1076,7 +1077,7 @@ write_hw_counter_info(std::ostream& os, const array_t<bool, N>& options,
auto _papi_events = tim::papi::available_events_info();
auto _rocm_events =
(gpu_count > 0) ? rocprofsys::rocprofiler::rocm_metrics() : hwcounter_info_t{};
(gpu_count > 0) ? rocprofsys::rocm::rocm_events() : hwcounter_info_t{};
if(alphabetical)
{
source/bin/rocprof-sys-avail/generate_config.cpp
+1 -1
Ver ficheiro
@@ -339,7 +339,7 @@ generate_config(std::string _config_file, const std::set<std::string>& _config_f
for(const auto* itr :
{ "ROCPROFSYS_CONFIG", "ROCPROFSYS_MODE", "ROCPROFSYS_TRACE",
"ROCPROFSYS_PROFILE", "ROCPROFSYS_USE_SAMPLING",
"ROCPROFSYS_USE_PROCESS_SAMPLING", "ROCPROFSYS_USE_ROCTRACER",
"ROCPROFSYS_USE_PROCESS_SAMPLING", "ROCPROFSYS_USE_ROCM",
"ROCPROFSYS_USE_ROCM_SMI", "ROCPROFSYS_USE_KOKKOSP",
"ROCPROFSYS_USE_OMPT", "ROCPROFSYS_USE", "ROCPROFSYS_OUTPUT" })
{
source/bin/rocprof-sys-avail/info_type.cpp
-2
Ver ficheiro
@@ -29,8 +29,6 @@
#include "library/components/fork_gotcha.hpp"
#include "library/components/mpi_gotcha.hpp"
#include "library/components/pthread_gotcha.hpp"
#include "library/components/rocprofiler.hpp"
#include "library/components/roctracer.hpp"
#include <timemory/components/definition.hpp>
#include <timemory/enum.h>
source/bin/rocprof-sys-causal/impl.cpp
-9
Ver ficheiro
@@ -752,10 +752,6 @@ parse_args(int argc, char** argv, std::vector<char*>& _env,
parser.end_group();
#if ROCPROFSYS_HIP_VERSION > 0 && ROCPROFSYS_HIP_VERSION < 50300
update_env(_env, "HSA_ENABLE_INTERRUPT", 0);
#endif
auto _inpv = std::vector<char*>{};
auto _outv = std::vector<char*>{};
bool _hash = false;
@@ -824,11 +820,6 @@ parse_args(int argc, char** argv, std::vector<char*>& _env,
add_default_env(_env, "ROCPROFSYS_USE_MPIP", true);
#endif
#if defined(ROCPROFSYS_USE_ROCTRACER) && ROCPROFSYS_USE_ROCTRACER > 0
add_default_env(_env, "ROCPROFSYS_ROCTRACER_HIP_API", true);
add_default_env(_env, "ROCPROFSYS_ROCTRACER_HSA_API", true);
#endif
#if defined(ROCPROFSYS_USE_RCCL) && ROCPROFSYS_USE_RCCL > 0
add_default_env(_env, "ROCPROFSYS_USE_RCCLP", true);
#endif
source/bin/rocprof-sys-instrument/CMakeLists.txt
+2
Ver ficheiro
@@ -35,6 +35,8 @@ target_link_libraries(
timemory::timemory-extensions
timemory::timemory-core)
add_target_flag_if_avail(rocprofiler-systems-instrument "-Wno-deprecated-declarations")
set_target_properties(
rocprofiler-systems-instrument
PROPERTIES BUILD_RPATH "\$ORIGIN:\$ORIGIN/../${CMAKE_INSTALL_LIBDIR}"
source/bin/rocprof-sys-instrument/internal_libs.cpp
+19 -7
Ver ficheiro
@@ -312,13 +312,25 @@ get_internal_basic_libs_impl()
"liblzma.so" };
// shared libraries used by rocprof-sys
const auto _omni_libs = strview_init_t{
"libstdc++.so.6", "libgotcha.so", "libunwind-coredump.so",
"libunwind-generic.so", "libunwind-ptrace.so", "libunwind-setjmp.so",
"libunwind.so", "libunwind-x86_64.so", "librocm_smi64.so",
"libroctx64.so", "librocmtools.so", "libroctracer64.so",
"librocprofiler64.so", "libpapi.so", "libpfm.so"
};
const auto _omni_libs = strview_init_t{ "libstdc++.so.6",
"libgotcha.so",
"libunwind-coredump.so",
"libunwind-generic.so",
"libunwind-ptrace.so",
"libunwind-setjmp.so",
"libunwind.so",
"libunwind-x86_64.so",
"librocm_smi64.so",
"libroctx64.so",
"librocmtools.so",
"libroctracer64.so",
"librocprofiler64.so",
"libpapi.so",
"libpfm.so",
"librocprofiler-register.so",
"librocprofiler-sdk.so",
"librocprofiler-sdk-roctx.so",
"libamd_smi.so" };
// shared libraries potentially used by timemory
const auto _3rdparty_libs = strview_init_t{ "libcaliper.so",
source/bin/rocprof-sys-instrument/rocprof-sys-instrument.cpp
+6 -4
Ver ficheiro
@@ -357,10 +357,12 @@ main(int argc, char** argv)
itr.find("rocprof-sys") != std::string::npos ||
itr.find("rocprofiler-systems") != std::string::npos ||
std::regex_search(
itr, std::regex{ "lib(dyninstAPI|stackwalk|pcontrol|patchAPI|parseAPI|"
"instructionAPI|symtabAPI|dynDwarf|common|dynElf|tbb|"
"tbbmalloc|tbbmalloc_proxy|gotcha|libunwind|roctracer|"
"hsa-runtime|amdhip|rocm_smi)\\.(so|a)" }))
itr, std::regex{
"lib(dyninstAPI|stackwalk|pcontrol|patchAPI|parseAPI|"
"instructionAPI|symtabAPI|dynDwarf|common|dynElf|tbb|tbbmalloc|"
"tbbmalloc_proxy|gotcha|libunwind|roctracer64|hsa-runtime|amdhip|"
"amd_comgr|rocm_smi64|rocprofiler64|rocprofiler-register|"
"rocprofiler-sdk|rocprofiler-sdk-roctx|amd_smi)\\.(so|a)" }))
{
if(!find(filepath::dirname(itr), lib_search_paths))
lib_search_paths.emplace_back(filepath::dirname(itr));
source/bin/rocprof-sys-sample/impl.cpp
+24 -72
Ver ficheiro
@@ -44,14 +44,6 @@
#include <unistd.h>
#include <vector>
#if !defined(ROCPROFSYS_USE_ROCTRACER)
# define ROCPROFSYS_USE_ROCTRACER 0
#endif
#if !defined(ROCPROFSYS_USE_ROCPROFILER)
# define ROCPROFSYS_USE_ROCPROFILER 0
#endif
namespace color = tim::log::color;
using namespace timemory::join;
using tim::get_env;
@@ -140,17 +132,6 @@ get_initial_environment()
update_env(_env, "ROCPROFSYS_USE_SAMPLING", (_mode != "causal"));
#if defined(ROCPROFSYS_USE_ROCTRACER) || defined(ROCPROFSYS_USE_ROCPROFILER)
update_env(_env, "HSA_TOOLS_LIB", _dl_libpath);
if(!getenv("HSA_TOOLS_REPORT_LOAD_FAILURE"))
update_env(_env, "HSA_TOOLS_REPORT_LOAD_FAILURE", "1");
#endif
#if defined(ROCPROFSYS_USE_ROCPROFILER)
update_env(_env, "ROCP_TOOL_LIB", _omni_libpath);
if(!getenv("ROCP_HSA_INTERCEPT")) update_env(_env, "ROCP_HSA_INTERCEPT", "1");
#endif
#if defined(ROCPROFSYS_USE_OMPT)
if(!getenv("OMP_TOOL_LIBRARIES"))
update_env(_env, "OMP_TOOL_LIBRARIES", _dl_libpath, UPD_APPEND);
@@ -357,14 +338,6 @@ parse_args(int argc, char** argv, std::vector<char*>& _env)
%{INDENT}% 0 avoid triggering the bug, potentially at the cost of reduced performance
%{INDENT}% 1 do not modify how ROCm is notified about kernel completion)";
auto _realtime_reqs = (get_env("HSA_ENABLE_INTERRUPT", std::string{}, false).empty())
? std::vector<std::string>{ "hsa-interrupt" }
: std::vector<std::string>{};
#if ROCPROFSYS_USE_ROCTRACER == 0 && ROCPROFSYS_USE_ROCPROFILER == 0
_realtime_reqs.clear();
#endif
const auto* _trace_policy_desc =
R"(Policy for new data when the buffer size limit is reached:
%{INDENT}%- discard : new data is ignored
@@ -720,7 +693,6 @@ parse_args(int argc, char** argv, std::vector<char*>& _env)
parser.add_argument({ "--realtime" }, _realtime_desc)
.min_count(0)
.required(std::move(_realtime_reqs))
.action([&](parser_t& p) {
auto _v = p.get<std::deque<std::string>>("realtime");
update_env(_env, "ROCPROFSYS_SAMPLING_REALTIME", true);
@@ -741,10 +713,20 @@ parse_args(int argc, char** argv, std::vector<char*>& _env)
}
});
std::set<std::string> _backend_choices = { "all", "kokkosp", "mpip",
"ompt", "rcclp", "rocm-smi",
"roctracer", "rocprofiler", "roctx",
"mutex-locks", "spin-locks", "rw-locks" };
std::set<std::string> _backend_choices = { "all",
"kokkosp",
"mpip",
"ompt",
"rcclp",
"rocm-smi",
"roctracer",
"rocprofiler",
"roctx",
"mutex-locks",
"spin-locks",
"rw-locks",
"rocprofiler-sdk",
"rocm" };
#if !defined(ROCPROFSYS_USE_MPI) && !defined(ROCPROFSYS_USE_MPI_HEADERS)
_backend_choices.erase("mpip");
@@ -758,17 +740,10 @@ parse_args(int argc, char** argv, std::vector<char*>& _env)
_backend_choices.erase("rcclp");
#endif
#if !defined(ROCPROFSYS_USE_ROCM_SMI)
#if !defined(ROCPROFSYS_USE_ROCM)
_backend_choices.erase("rocm");
_backend_choices.erase("rocm-smi");
#endif
#if !defined(ROCPROFSYS_USE_ROCTRACER)
_backend_choices.erase("roctracer");
_backend_choices.erase("roctx");
#endif
#if !defined(ROCPROFSYS_USE_ROCPROFILER)
_backend_choices.erase("rocprofiler");
_backend_choices.erase("rocprofiler-sdk");
#endif
parser.start_group("BACKEND OPTIONS",
@@ -784,11 +759,9 @@ parse_args(int argc, char** argv, std::vector<char*>& _env)
_update("ROCPROFSYS_USE_KOKKOSP", _v.count("kokkosp") > 0);
_update("ROCPROFSYS_USE_MPIP", _v.count("mpip") > 0);
_update("ROCPROFSYS_USE_OMPT", _v.count("ompt") > 0);
_update("ROCPROFSYS_USE_ROCM", _v.count("rocm") > 0);
_update("ROCPROFSYS_USE_RCCLP", _v.count("rcclp") > 0);
_update("ROCPROFSYS_USE_ROCTX", _v.count("roctx") > 0);
_update("ROCPROFSYS_USE_ROCM_SMI", _v.count("rocm-smi") > 0);
_update("ROCPROFSYS_USE_ROCTRACER", _v.count("roctracer") > 0);
_update("ROCPROFSYS_USE_ROCPROFILER", _v.count("rocprofiler") > 0);
_update("ROCPROFSYS_TRACE_THREAD_LOCKS", _v.count("mutex-locks") > 0);
_update("ROCPROFSYS_TRACE_THREAD_RW_LOCKS", _v.count("rw-locks") > 0);
_update("ROCPROFSYS_TRACE_THREAD_SPIN_LOCKS", _v.count("spin-locks") > 0);
@@ -810,27 +783,18 @@ parse_args(int argc, char** argv, std::vector<char*>& _env)
_update("ROCPROFSYS_USE_KOKKOSP", _v.count("kokkosp") > 0);
_update("ROCPROFSYS_USE_MPIP", _v.count("mpip") > 0);
_update("ROCPROFSYS_USE_OMPT", _v.count("ompt") > 0);
_update("ROCPROFSYS_USE_ROCM", _v.count("rocm") > 0);
_update("ROCPROFSYS_USE_RCCLP", _v.count("rcclp") > 0);
_update("ROCPROFSYS_USE_ROCTX", _v.count("roctx") > 0);
_update("ROCPROFSYS_USE_ROCM_SMI", _v.count("rocm-smi") > 0);
_update("ROCPROFSYS_USE_ROCTRACER", _v.count("roctracer") > 0);
_update("ROCPROFSYS_USE_ROCPROFILER", _v.count("rocprofiler") > 0);
_update("ROCPROFSYS_TRACE_THREAD_LOCKS", _v.count("mutex-locks") > 0);
_update("ROCPROFSYS_TRACE_THREAD_RW_LOCKS", _v.count("rw-locks") > 0);
_update("ROCPROFSYS_TRACE_THREAD_SPIN_LOCKS", _v.count("spin-locks") > 0);
if(_v.count("all") > 0 ||
(_v.count("roctracer") > 0 && _v.count("rocprofiler") > 0))
{
remove_env(_env, "HSA_TOOLS_LIB");
remove_env(_env, "HSA_TOOLS_REPORT_LOAD_FAILURE");
}
if(_v.count("all") > 0 || _v.count("rocprofiler") > 0)
{
remove_env(_env, "ROCP_TOOL_LIB");
remove_env(_env, "ROCP_HSA_INTERCEPT");
}
// if(_v.count("all") > 0 || _v.count("rocprofiler") > 0)
// {
// remove_env(_env, "ROCP_TOOL_LIB");
// remove_env(_env, "ROCP_HSA_INTERCEPT");
// }
if(_v.count("all") > 0 || _v.count("ompt") > 0)
remove_env(_env, "OMP_TOOL_LIBRARIES");
@@ -850,18 +814,6 @@ parse_args(int argc, char** argv, std::vector<char*>& _env)
update_env(_env, "ROCPROFSYS_PAPI_EVENTS", _events);
});
#if defined(ROCPROFSYS_USE_ROCPROFILER)
parser
.add_argument({ "-G", "--gpu-events" },
"Set the GPU hardware counter events to record (ref: "
"`rocprof-sys-avail -H -c GPU`)")
.action([&](parser_t& p) {
auto _events =
join(array_config{ "," }, p.get<std::vector<std::string>>("gpu-events"));
update_env(_env, "ROCPROFSYS_ROCM_EVENTS", _events);
});
#endif
parser.start_group("MISCELLANEOUS OPTIONS", "");
parser
.add_argument({ "-i", "--inlines" },